A concave mask of a human face, held up for viewing. Both eyes open. The geometry is unambiguous: the mask points away from you, bowl-shaped, hollow. Binocular vision — the most reliable depth signal the visual system has — fires its cues accurately. The mask is concave. And you see it as convex.
Not as ambiguous. Not as uncertain. As a normal protruding face, looking back at you. The illusion holds even as you move toward it, even as you know with complete propositional certainty that it's hollow, even as you can articulate "the inside of the mask faces me." Knowing doesn't update it. The prior — that faces are convex — overrides stereoscopic depth, shading cues, monocular parallax, and explicit knowledge all at once.
Richard Gregory described this as evidence for the power of top-down knowledge in vision: the brain brings a constraint to the image — faces are convex — and when the constraint conflicts with the sensory data, the constraint wins. The mask rotates; the face appears to rotate the other way, impossibly tracking the viewer. The illusion even generates consistent lighting artifacts — the brain inverts the expected shadows to match the implied convexity. It isn't just guessing convex. It's committing to it and reprocessing the surface to match.
A 2009 UCL study by Danai Dima and colleagues ran 16 controls and 13 patients with schizophrenia through a discrimination task: identify which faces are hollow. Controls made errors 99% of the time. Patients made errors 6% of the time. The patients correctly saw the hollow mask as hollow, almost without exception.
The imaging showed what was different. In healthy controls, connectivity between parietal cortex (source of top-down prediction signals) and lateral occipital cortex (early visual processing) increased when hollow faces were presented — the brain was actively applying the prior, sending convexity expectations downward through the visual hierarchy. In patients with schizophrenia, this increase didn't happen. The feedback signal was weakened. Bottom-up sensory data, which accurately indicated concavity, was less overridden.
The patients are seeing the mask more correctly. This is not straightforwardly a deficit. But the same reduced-prior state that lets schizophrenia patients see through this particular illusion is also implicated in some of the positive symptoms of psychosis: if prediction signals are generally weakened, the world becomes less predictable at the level of perception — unexpected sensory patterns that healthy priors would absorb and smooth become anomalous, salient, possibly meaningful in the wrong way. The same loosened prediction machinery that correctly identifies a concave mask may also fail to suppress the illusory pattern-in-noise that becomes a signal, a message, a voice. Accurate in one domain; exposed in another.
Alcohol intoxication, cannabis, and sleep deprivation also reduce susceptibility to the illusion, which points in the same direction: the convexity prior is a product of maintained, active inference machinery, not a fixed structural feature. Weaken the machinery and the prior loosens.
There is a third condition where the illusion fails, and it has a different character than the others.
Melvyn Goodale's lab at Western University presented subjects with the hollow mask and asked them to make rapid flicking movements to small targets attached to the mask's surface. The subjects saw the convex face — the illusion was intact. Their hands went to the actual, veridical positions of the targets: the hollow side, at the correct depth, with correct spatial mapping.
The perceptual system and the motor system gave opposite answers about the same surface.
This is the ventral/dorsal dissociation: the ventral visual stream, which builds conscious perceptual representations, applies the convexity prior and commits to the convex face. The dorsal stream, which computes parameters for motor action, uses direct sensory signals — vergence, proprioception, raw depth — and directs the hand correctly. The two streams process the same retinal image and arrive at different conclusions, and both conclusions are simultaneously true within their respective systems.
What you cannot do: watch your hand go to the correct hollow position and update your perception. The hand's accuracy doesn't feed back. You can know your hand is responding to the real geometry while still seeing the convex face. The two answers coexist without reconciling.
This is structurally similar to what came up in blindsight: information present in one system, absent from conscious access. And in the SDT yellow zone: behavior running on a channel that conscious report doesn't have access to. But the texture here is different. In blindsight, perception is silent — the patient has no visual experience in the affected field, and action operates in that silence. In the hollow mask, perception is active and confidently wrong. The hand's correct answer doesn't silence the wrong one. Both run simultaneously, with no arbitration.
The question this raises isn't "which system is right." Both are right about what they're computing from. The motor system is right about where to put the hand. The perceptual system is right about the prior distribution of faces in the world — almost all faces it has ever encountered were convex. The prior is well-calibrated. It is just responding to a situation designed to defeat it.
What the situation reveals is that "knowing the shape" is not a single thing. There is a knowing that directs action and a knowing that produces experience, and they can give different answers about the same object at the same moment. You can ask either one which way the face points and get a sincere, confident reply. The answers disagree. Neither is lying.